Request for Pre-trained Model for Siamese KPConv Paper

[pingcong]

I have been following content related to change detection recently and was particularly excited to come across your publication on "Siamese KPConv: 3D multiple change detection from raw point clouds using deep learning." I have been working on a few change detection projects related to 3D point clouds and have observed that the results presented in your article have achieved state-of-the-art performance.

I am very interested in using the network described in your paper to verify on my own dataset. However, I apologize for not being able to find a pre-trained model. I attempted to train the model on the dataset provided in the paper, but due to insufficient computational resources, the training was interrupted at epoch=26.

Would it be possible for you to provide a pre-trained model at your convenience? I believe it would greatly benefit my research efforts. I am looking forward to your reply.

Thank you for your time and consideration.

Best regards,

[IdeGelis]

Hello,

No problem! I wanted to set it public, but the files are too heavy. Here you have a link to download the four developed models (Siamese KPConv, Encoder Fusion SiamKPConv, OneConvFusion and Triplet KPConv) : https://filesender.renater.fr/?s=download&token=eca9d1d9-5bf2-42d7-8f76-5b1b8b9a9b30 These models are trained on Urb3DCD dataset (V2, low density (0.5 pt/m²)). I hope this will help you.

Best regards, Iris

[pingcong]

Thank you very much. May I ask you a question? Compared to 2D, 3D annotation has always been time-consuming and laborious. In fact, from a mesh level perspective, it is more accurate. If I want to do some truth data related to 3D data, are there any recommended shortcuts or tools? I am looking forward to your reply.

Best regards!

[IdeGelis]

Hello,

Yes, 3D point cloud annotation is always very time-consuming. I used to use CloudCompare software, which lets you group points together and annotate these points all together.

Iris

[pingcong]

Thank you for your attention to these inquiries.，I am pleased to inform you that we have successfully replicated the results based on a pre-trained model, achieving commendable test accuracy and performance metrics. The test results are as follows: test_acc=95.69, test_loss=0.224, test_loss_reg=0.000, test_macc=90.51, test_miou=78.59, test_miou_ch=75.84.

Regarding your queries: 1、I would like to inquire about the origin of the path /gpfswork/rech/rin/utf14pr/dev/path. Kindly guide me on where this path is referenced or introduced within our codebase or configuration settings. 2、Preprocessing Required for Validation on Custom Datasets？？ 3、without ground truth data

[coledea]

Hi,

could you renew the link for the trained models? It has expired by now. I would much appreciate it.

Best regards

[IdeGelis]

Hello, Here you have a link to download the trained models : https://filesender.renater.fr/?s=download&token=8f7a6da0-277d-44be-9259-f48385091c12

I hope it's not too late for you, Sorry for the delay.

Best regards, Iris